[Devel] [PATCH] binfmt_misc: fix mount after umount in CT

[Alexander Mikhalitsyn]

On Mon, 18 Oct 2021 19:51:49 +0300
Nikita Yushchenko <nikita.yushchenko at virtuozzo.com> wrote:

> The assumption that bm_fill_super() is not called for the second time
> for CT is wrong: umount operation clears sb->s_root, which causes
> vfs_get_super() to call fill_super again on the next mount.
> 
> Make bm_fill_super() handle multiple-calls corrently:
> - initialize bm_data and set ve->binfmt_misc only if it is not done
>   before,
> - delay desctruction of it up to CT destruction.
> 
> https://jira.sw.ru/browse/PSBM-133968
> Fixes: 8250ff41d190 ("ve/fs/binfmt: clean bm_data reference from ve on err path")
> Signed-off-by: Nikita Yushchenko <nikita.yushchenko at virtuozzo.com>
> ---
>  fs/binfmt_misc.c | 37 +++++++++++++------------------------
>  1 file changed, 13 insertions(+), 24 deletions(-)
> 
> diff --git a/fs/binfmt_misc.c b/fs/binfmt_misc.c
> index 0946e7e6caa5..5f4e90c1ade2 100644
> --- a/fs/binfmt_misc.c
> +++ b/fs/binfmt_misc.c
> @@ -853,38 +853,25 @@ static int bm_fill_super(struct super_block *sb, struct fs_context *fc)
>  	};
>  
>  	struct ve_struct *ve = get_exec_env();
> -	struct binfmt_misc *bm_data;
> +	struct binfmt_misc *bm_data = ve->binfmt_misc;
>  
> -	/*
> -	 * bm_get_tree()
> -	 *  get_tree_keyed(fc, bm_fill_super, get_ve(ve))
> -	 *   fc->s_fs_info = current VE
> -	 *   vfs_get_super(fc, vfs_get_keyed_super, bm_fill_super)
> -	 *    sb = sget_fc(fc, test, set_anon_super_fc)
> -	 *    if (!sb->s_root) {
> -	 *		err = bm_fill_super(sb, fc);
> -	 *
> -	 * => we should never get here with initialized ve->binfmt_misc.
> -	 */
> -	if (WARN_ON_ONCE(ve->binfmt_misc))
> -		return -EEXIST;
> +	if (!bm_data) {
> +		bm_data = kzalloc(sizeof(struct binfmt_misc), GFP_KERNEL);
> +		if (!bm_data)
> +			return -ENOMEM;
>  
> -	bm_data = kzalloc(sizeof(struct binfmt_misc), GFP_KERNEL);
> -	if (!bm_data)
> -		return -ENOMEM;
> +		INIT_LIST_HEAD(&bm_data->entries);
> +		rwlock_init(&bm_data->entries_lock);
>  
> -	INIT_LIST_HEAD(&bm_data->entries);
> -	rwlock_init(&bm_data->entries_lock);
> +		ve->binfmt_misc = bm_data;

Isn't it better to move ve->binfmt_misc assignment to the
end of function where we know that all operations was successful?

> +		/* this will be cleared by ve_binfmt_fini() */
> +	}
>  
>  	err = simple_fill_super(sb, BINFMTFS_MAGIC, bm_files);
> -	if (err) {
> -		kfree(bm_data);

If we have ve->binfmt_misc assignment in the upper part of code, then
we need to do ve->binfmt_misc = NULL here.

> +	if (err)
>  		return err;
> -	}
>  
>  	sb->s_op = &s_ops;
> -
> -	ve->binfmt_misc = bm_data;
see above

>  	bm_data->enabled = 1;
>  
>  	return 0;
> @@ -971,6 +958,8 @@ static void ve_binfmt_fini(void *data)
>  	while (!list_empty(&bm_data->entries))
>  		kill_node(bm_data, list_first_entry(
>  			&bm_data->entries, Node, list));
> +
> +	kfree(bm_data);

We have kfree in ve_destroy (kernel/ve/ve.c) already.

>  }
>  
>  static struct ve_hook ve_binfmt_hook = {
> -- 
> 2.30.2
>